Unsupervised large-cone-angle artifact suppression for CBCT reconstruction via neural radiance fields.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Science

Background:

• Cone-beam computed tomography (CBCT) is crucial for clinical diagnostics.
• Circular trajectories in CBCT acquisition lead to significant large-cone-angle artifacts (LCAA) due to null space deficiency.
• Current artifact reduction methods often require extensive prior information or paired data, hindering clinical use.

Purpose of the Study:

• To develop an unsupervised framework for suppressing LCAA in CBCT directly from projection data.
• To overcome limitations of existing methods that require extensive prior information or paired data.
• To improve the clinical applicability of CBCT by enhancing image quality.

Main Methods:

• Proposed an unsupervised Neural Radiance Fields framework (LCAA-NeRF).
• Introduced an axial-aware anisotropic scaling hash encoding (A³Hash) for enhanced representation of large cone angles.
• Implemented a path-length-adaptive ray sampling (PLARS) strategy for dynamic feature capture.
• Incorporated a stochastic structural similarity (S3IM) loss for geometric consistency.

Main Results:

• Successfully suppressed LCAA in CBCT reconstruction using the LCAA-NeRF framework.
• Demonstrated the effectiveness of A³Hash and PLARS in handling large cone angles and varying ray paths.
• Validated the method's superiority and robustness on both simulated and real-world CBCT datasets.
• Achieved improved geometric consistency through the S3IM loss.

Conclusions:

• LCAA-NeRF offers an effective unsupervised solution for LCAA reduction in CBCT.
• The proposed A³Hash encoding and PLARS sampling strategies enhance CBCT image reconstruction quality.
• This method shows significant potential for improving clinical CBCT imaging applications.